Learning Deep Spatial Lung Features by 3D Convolutional Neural Network for Early Cancer Detection

Abstract

Accurate early lung cancer detection is essential towards precision oncology and would effectively improve the patients' survival rate. In this work, we explore the lung cancer early detection capacity by learning from deep spatial lung features. A 3D CNN network architecture is constructed with segmented CT lung volumes as training and testing samples. The new model extracts and projects 3D features to the following hidden layers, which preserves the temporal relations between neighboring CT slices. The well-built 3D CNN model consists of 11 layers which generates 12,544 neurons and 16 million parameters classifying whether the patient is diagnosed as cancer or not. ReLU nonlinearity and Sigmoid function are used as activation and classification methods. The model achieves a prediction accuracy of 87.5% where only the biomedical images themselves are used as the input dataset. The model's lowest error rate reaches 12.5% that improves the traditional AlexNet architecture by 2.8%.